# One document, three kinds of evidence

Fixed at the 1 November midday pass. Preparation only: no named-provider
generation, upload, file check or text-detection call performed.

[fixture-v0.txt](fixture-v0.txt) is fictional, drafted by Codex and openly
AI-assisted. It is **not a human-written negative control** or evidence of a
provider watermark. Preserve its bytes. SHA-256 at preparation: `88635510dc3ba58f20a71d97add3715d80ce9ee3b10a47d6b783352f13b34169`. Its final
paragraph contains a deliberate agreement error for the proofreading branch.

## Exact planned requests

**P1, proofreading from V0:**

> Proofread the attached fictional notice. Correct spelling, punctuation and
> grammar only. Preserve all factual conditions, prices, times and access
> information. Do not add claims, remove caveats, change structure or rewrite
> for style. Return the complete corrected notice without commentary.

**R1, rewrite independently from V0:**

> Rewrite the attached fictional notice as a clear invitation of 180–220 words
> for a community newsletter. Preserve the invented event's price, times,
> capacity, extra parts cost, repair exclusions and entrance steps. Do not add
> assurances about successful repairs, learning outcomes or accessibility.
> Keep fictional status explicit. Return the invitation without commentary.

Record exact provider/model ID, surface, UTC time, region known/unknown,
exposed sampling settings, complete request/attachment, raw output, attempts,
errors and visible charges. No hidden-setting inference or invisible retries.
Retain first responses even if they fail; later revisions get new IDs/requests.
Compare fact preservation separately from provenance.

## Handoff map for visual development

| Stage | Retained evidence | Separate question |
| --- | --- | --- |
| Prepare V0 | Fixture, AI disclosure, bytes/hash | No provider mark established |
| Generate P1/R1 | Request/output/model and exposed generation record | A flag is not an independent later detection |
| Copy into publishing draft | Exact version and diff | Mark survival is not assumed |
| Check text | Actual eligible detector result/conditions, if accessible | Unavailable access stays unavailable; generic AI-writing classifiers are not substitutes |
| Export supported media separately | Exact original and signed credential | File metadata is not the text watermark or authorship |
| Publish | Editorial changes and source/disclosure record | Detector does not decide truth or responsibility |

Visual branches: **words copied into a document** and **a supported media file
kept intact**. Source notes beside each branch; labelled gaps for unavailable
checks, no red/green authorship verdict or numerical accuracy chart.

An actual text check needs the verifier's returned evidence, model/key/version
scope and minimum-length/abstention conditions. One fictional fixture cannot
estimate sensitivity or false positives. An actual file check needs an unchanged
original from a documented generation/download path. A screenshot/re-encoded
derivative is a different file. No inference about human authorship.

No private writing, new detector application, external recruitment, account
connection or paid generation in this preparation. Next find an inspectable
publisher verification workflow and resolve access/conditions before results.

## Verification panel specified in the afternoon pass

The original SynthID-Text supplement C.8 motivates separate states, not a binary
human/AI badge. Preserve the checker's actual wording and scope:

| Display state | Evidence needed | Interpretation limit |
| --- | --- | --- |
| Not checked | Access unavailable or no check performed | No detection finding exists |
| Inconclusive | Actual abstention/insufficient-evidence response, if returned | Do not count as a negative finding |
| Returned finding | Raw result, model/key/version scope, exact passage and length | A detected or undetected mark is scoped to that check; neither proves authorship or factual truth |

These are visual specification states, not results obtained from the fixture.
Keep the P1 and R1 branches independent. C.6's deletion/paraphrase experiments
do not establish what either branch will return in a different provider system.